Voices
Mathematician Andreas Thom asks whether OpenAI's non-sofic group proof drew on his private ChatGPT sessions
TU Dresden group theorist Andreas Thom posted a three-part Mathstodon thread saying OpenAI's model built directly on his 2019 paper with Gábor Kun to construct the first non-sofic group, after he and a colleague spent months working related problems inside ChatGPT. He asked whether those chats entered training data or were reachable during the proof; he says OpenAI's Mark Sellke replied only that 'that did not happen,' and that anonymization does not protect unpublished mathematical ideas. He is the second mathematician this week to raise the question, which hit 825 points on HN, and any researcher pasting unpublished work into a chat product now has a concrete reason to check data-use settings.
↳ Follow the thread